Proceedings of the Annual Meeting of the Cognitive Science Society

Emergent communication in artificial agents has been studiedto understand language evolution, as well as to develop artifi-cial systems that learn to communicate with humans. We showthat agents performing a cooperative navigation task in variousgridworld environments learn an interpretable communicationprotocol that enables them to efficiently, and in many cases,optimally, solve the task. An analysis of the agents’ policiesreveals that emergent signals spatially cluster the state space,with signals referring to specific locations and spatial direc-tions such as left, up, or upper left room. Using populationsof agents, we show that the emergent protocol has basic com-positional structure, thus exhibiting a core property of naturallanguage.